You’ve been spending a lot of time with companies discussing how to prevent customer churn. Any consistent theme you are seeing across these discussions?
First, I think we need to re-frame the issue. Instead of focusing on “customer churn” let’s turn that problem on its head and instead focus on “customer success”. And what I mean by that is so often we are late to the party. Companies are often identifying customers who are likely to churn when it’s too late. For example, order cadence is not a leading indicator. It’s more of a lagging indicator of a problem that has already manifested itself. In many cases, the best signal for customer churn is further back in the customer lifecycle at the point of acquisition and on-boarding.
Surprisingly many companies I’ve talked to can’t tell me what an acceptable level of churn is. Not all customer attrition is regrettable. It might be ok for certain customers to churn when the cost-to-serve is high and the margins are low. And that assumes we are acquiring net-new customers at an appropriate velocity and volume to more than compensate for our lost business. But you need to have those basic analytics in place to benchmark what an acceptable level of churn is. That is foundational.
I’m also blown away by service-based businesses that fail to marry up service utilization data with their CRM data. This screams “low hanging fruit” to me. If you’re a customer that hasn’t activated your service – or your utilization of that service is dismal, then you might have a problem. It seems intuitive enough, but often this data is locked away in different systems and nobody has gone through the trouble to surface that information at relevant moments in the workflow.
Today we use logistic regression models to predict the probability of a customer to churn. These machine learning models are trained on historical data where we can see or infer examples of customers that have either churned or not churned. Once we have a probability score for a given customer we can then take a multitude of different actions to influence that probability and hopefully change the outcome. We also use natural language processing models to discern a customer’s sentiment. We can feed these models large amounts of unstructured data, including call recordings or web chats, and start to isolate emerging themes. Do these customers feel good, bad or indifferent? This classification will often be input into a logistic regression-based customer churn model. So, in essence, you are starting to chain together multiple models that can help isolate customers that are likely to churn.
It’s a pretty broad spectrum. In some cases, a specialized team or a “retention desk” is given a simple call down list to reach out to customers that have a high propensity of churning. As we step up the maturity curve, we see “next best actions” surfaced to service agents that are part of an elaborate call script. And as we progress further up the maturity curve we see companies embedding churn predictions into their overall “Universal Agent” strategy. As a customer service agent, I can leverage these insights and recommended actions to pivot a customer complaint into a selling motion. Or maybe I’m onboarding that customer onto a new marketing journey. I love the notion of empowering customer service agents with insights that allow an organization to start blurring the lines between sales, service, and marketing. The best time to have a selling conversation is right after you’ve solved a problem.